We cannot test the hypothesis yet, so we tested the scoreboard that will judge it.

PIPELINE VALIDATION UNDER SYNTHETIC WORLDS · NO HARDWARE · PHYSICAL EVIDENCE NONE

COMPUTATIONAL RESULT (SYNTHETIC DATA, ASSUMED NOISE MODEL)

Every number on this page is generated from assumed distributions we wrote ourselves. A high classification rate says our scoreboard is hard to fool by the failure modes we imagined. It says nothing whatsoever about whether the physical effect exists.

Phase 71b · third ruler promoted

The two-ruler test is necessary but not sufficient.

THE TWO-RULER TEST IS NECESSARY BUT NOT SUFFICIENT. Three ordinary mechanisms — adaptive caching, a hidden central scheduler, and a precomputed shortcut table — pass both scaling rulers in roughly 95% of synthetic run sets. A third ruler, mechanism locality and auditability, is now a hard gate.

Evil twins

Three ordinary mechanisms wearing our result as a costume

Each mimic is a mechanism nobody would call exotic. We chose their target means specifically so that both frozen scaling rulers are satisfied. Over 500 synthetic run sets each, they pass the old rule roughly 95% of the time.

Third ruler · mechanism locality and auditability

Ten conditions, frozen before hardware exists

A would-be analogue candidate must now prove not only that the two numbers move the right way, but that the mechanism producing them is genuinely local and fully auditable. These conditions are fixed now, while we still have no data to bias them.

Three-ruler bench test · verdict engine

Only all three rulers plus the quality gates can yield ANALOGUE HARDWARE EFFECT CANDIDATE. Passing rulers 1 and 2 while failing the audit is now explicitly labelled CONVENTIONAL / UNAUDITABLE MECHANISM — the evil twins live there.

In plain language

Two scoreboards can be fooled by a magician who already knows the maze. So we added a third referee who watches every move and checks that nobody secretly looks at the whole map.

Frozen verdict engine

A pure function. It takes the frozen metric table and returns one of three labels. No analyst input is accepted at verdict time.

These thresholds are EXPERIMENT-DESIGN CHOICES and preregistered targets. They are not laws of nature, and no physical claim follows from meeting them.

We generated 1000 blinded synthetic datasets from each of three worlds where we already know the right answer, at N = 16, 32, 64, with 60 trials per size, and fed them to the frozen verdict engine. This measures how hard our scoreboard is to fool — nothing else.

These are PIPELINE-VALIDATION rates under distributions we assumed. They are not evidence for the physical hypothesis and can never become evidence for it.

Why the adjacent-monotonic rule was rejected

The damaging cell is the conventional-switching row: the brittle rule hands an ordinary programmable switch the analogue-candidate label whenever three noisy medians happen to fall in order. That is exactly the mistake the whole programme exists to avoid.

We vary the assumed latency noise, sparsity noise, trial count and a small systematic timing offset, and record the classification rate for each world. The rule is unchanged throughout — only the world's nastiness changes.

Read the failure modes honestly: the null world is never mislabelled, but a systematic timing offset is the engine's soft spot — it drags the conventional-switching world across the latency line and inflates apparent advantage. Clock calibration is therefore a first-class instrumentation requirement, not an afterthought. Small trial counts hurt the switching row too, which is why the run plan asks for ≥ 20 valid runs per condition.

Negative control · illegal controller write

One illegal controller write is injected into an otherwise perfect would-be analogue pass. The engine must and does drop it to FAIL / NO EFFECT. There is no severity scale and no judgement call: a single global read or write outside a node's allowed local neighbourhood invalidates the run set.

Data-integrity rules

• No manual exclusions. Ever. A run is dropped only by a preregistered hardware-fault code.
• No smoothing-dependent metric. Medians over raw samples, fixed thresholds, fixed dwell windows.
• Raw timestamps preserved verbatim alongside every derived number.
• Exclusion codes are fixed in advance: instrumentation failure, saturation / clipping, missing samples, clock-sync loss, explicit locality violation.
• Every excluded run ID and its code is published with the valid runs.
• The verdict engine is a pure function of the frozen metric table. No analyst input is accepted at verdict time.

Status reconciliation · stricter record wins

Where a later summary and an earlier recorded experiment disagree, the site keeps the STRICTER recorded evidence. Conversational claims that the multi-user, fault-tolerance and lifecycle gates were fully passed are superseded by the recorded Phase 66-69 verdicts, which are conditional. The hardware programme may proceed as an ANALOGUE EXPERIMENT DESIGN regardless — a bench test of a cost-and-sparsity signature does not require the toy architecture gate to be closed.